A LONGITUDINAL STUDY OF THE INFLUENCE OF ALLIANCE
NETWORK STRUCTURE AND COMPOSITION ON FIRM
EXPLORATORY INNOVATION
COREY C. PHELPSHEC Paris
This study examines the influence of the structure and composition of a firm’s alliance network on its exploratory innovation—innovation embodying knowledge that is novel relative to the firm’s extant knowledge. A longitudinal investigation of 77 telecommunications equipment manufacturers indicated that the technological diver-sity of a firm’s alliance partners increases its exploratory innovation. Further, network density among a firm’s alliance partners strengthens the influence of diversity. These results suggest the benefits of network “closure” (wherein a firm’s partners are also partners) and access to diverse information can coexist in an alliance network and that these combined benefits increase exploratory innovation.
A core area of research on strategic alliances con-cerns their influence on firm performance (Gulati, 1998). Within this domain of inquiry, researchers have often characterized alliances as wellsprings of innovation and new capabilities (e.g., Hamel, 1991; Leonard-Barton, 1995). Many studies have shown the alliance networks in which firms are embedded can enhance firm learning and innovation (e.g., Ahuja, 2000; Shan, Walker, & Kogut, 1994; Smith-Doerr et al., 1999; Soh, 2003). Despite this evi-dence, substantial opportunity exists to expand un-derstanding of how and under what conditions alliance networks influence firm innovation. A re-view of nearly 40 years of research published in 12 leading management and social science journals (Phelps, Heidl, & Wadhwa, 2010) showed the
liter-ature on alliances and firm innovation is limited in at least four important respects.
First, although some research has examined the influence of alliance network structure on firm in-novation, the composition of firms in these net-works has received little attention. Network struc-ture refers to the pattern of relationships that exist among a set of actors, and network composition refers to the types of actors in a network character-ized in terms of their stable traits, features, or re-source endowments (Wasserman & Faust, 1994). Recent research has recognized that alliance net-work studies have largely ignored netnet-work composi-tion and has called for more attencomposi-tion in network research to the heterogeneity of the resources of firms in networks (Lavie, 2006; Maurer & Ebers, 2006). The few studies that have examined both structure and composition have focused on the depth of partner technological resources and found that they improve a firm’s innovation performance (Baum, Calabrese, & Silverman, 2000; Stuart, 2000). Although dyad-level research has examined the influence of technological differences between partners on firm innovation (Sampson, 2007), research has largely overlooked the influence of network-level technological diversity— the technological differences between a firm and its partners and among the partners. Such compositional diversity is relevant to a current debate in the social network and alliance literatures.
Second, research has yielded conflicting results about the influence of alliance network structure on firm innovation. Research that examines the influence of social networks on creativity and in-novation has stressed the benefits actors derive from network structure and explored how these benefits, or “structural social capital” (Nahapiet &
This article is based on my doctoral dissertation, which received the Academy of Management’s TIM Division’s Best Dissertation Award and the State Farm Companies Foundation Dissertation Award. I thank my dissertation committee members, Raghu Garud, Theresa Lant, and J. Myles Shaver, for their advice and guidance. I am grateful for comments on drafts of this article by Sanjay Jain, Mel-issa Schilling, J. Myles Shaver, Kevin Steensma, Kate Stovel, Anu Wadhwa, and Mina Yoo. I thank Simon Rodan for his help with computing the diversity measure used here. I also thank Associate Editor Chet Miller for his me-ticulous feedback and thank the three anonymous review-ers for their comments, all of which greatly contributed to the improvement of the article. I acknowledge financial support from the State Farm Companies Foundation and the Berkley Center for Entrepreneurship at the Stern School of Business, NYU. Finally, I am grateful for the extraordi-nary database development and programming skills of Ralph Heidl and Tim Nali. All errors are my responsibility.
890
Copyright of the Academy of Management, all rights reserved. Contents may not be copied, emailed, posted to a listserv, or otherwise transmitted without the copyright holder’s express written permission. Users may print, download or email articles for individual use only.
Ghoshal, 1998), influence knowledge creation. In particular, the configuration of an actor’s set of direct ties (i.e., the actor’s “egocentric network structure”) has received some attention. This re-search has focused on triadic closure (i.e., whether an actor’s partners are partners), but two competing perspectives exist, each with different causal mech-anisms linking network structure to innovation. The argument of one view is that disconnected networks increase creativity and innovation be-cause they provide actors with timely access to
diverse information (Burt, 1992, 2004).1 An
alter-native view suggests dense networks, in which tri-ads are closed and “structural holes” (unconnected partners) are absent, provide social capital because such structures generate trust, reciprocity norms, and a shared identity, which increase cooperation and knowledge sharing (Coleman, 1988; Portes, 1998). Research has found support for both views, yielding conflicting results. Although studies have found structural holes in a firm’s network enhance its knowledge creation (Hargadon & Sutton, 1997; McEvily & Zaheer, 1999), other research has sug-gested that network closure improves knowledge transfer and innovation (Ahuja, 2000; Dyer & No-beoka, 2000; Schilling & Phelps, 2007).
One plausible reason for these conflicting results is that most studies have examined the influence of network structure and largely overlooked network composition. An examination of network composi-tion may help resolve these conflicting results and lead to a better understanding of how alliance net-works influence firm innovation. Another possible explanation is that different studies have examined different types of ties, different institutional con-texts, and different outcome variables. It is unlikely a particular network structure is universally bene-ficial (see Adler & Kwon, 2002). Research has sug-gested that the value of open versus closed net-works for innovation and creativity is contingent on the type of task (Hansen, 1999), type of tie (Ahuja, 2000), and particular institutional environ-ment (Owen-Smith & Powell, 2004). In contrast, network structure may act as a contingency vari-able and moderate the influence of network com-position on firm innovation. Moreover, this effect may depend on the type of learning and innovation actors pursue. Both of these contingencies have
been largely unexplored in prior research and are examined in this study.
A third limitation of research on alliances and firm innovation concerns an often-used, yet largely unexamined, assumption about the benefits of structural holes. Although a principal benefit at-tributed to structural holes is timely access to di-verse information (Burt, 1992), structural holes are neither a necessary nor a sufficient condition for such access (Reagans, Zuckerman, & McEvily, 2004). The informational benefits contacts provide can be directly observed by examining the extent to which they specialize in different domains of knowledge (Reagans & McEvily, 2003; Rodan & Galunic, 2004). Observing differences in competen-cies also allows a finer-grained measure of diversity than simply counting structural holes. Because competencies are stable and durable properties of firms (Patel & Pavitt, 1997), they are a composi-tional variable. Ties to partners with dissimilar knowledge stocks provide a firm with access to diverse information and know-how, independent of the structure of its local network. Thus, the so-cial control benefits of network closure and access to diverse information and know-how can coexist. Research on interfirm alliances (Ahuja, 2000) and interpersonal networks (Rodan & Galunic, 2004) has shown network density and knowledge diver-sity are empirically distinct. However, alliance re-search has not examined the independent and in-teractive effects of network structure and network knowledge diversity on firm innovation.
A final limitation of research on alliance net-works and firm innovation is that it largely ignores the novelty of the knowledge created and embodied in the innovations measured. Instead, studies have focused on the amount of innovation indicated by survey items or counts of new products and pat-ents. This approach implicitly rests on the as-sumption that innovations are similar in their knowledge content. Although research has sug-gested firms typically search for innovative solu-tions to problems in the domains of their existing expertise (“local search”) and produce “exploi-tive” innovations that represent incremental im-provements to their prior innovative efforts (e.g., Dosi, 1988; Martin & Mitchell, 1998), some re-search has shown firms vary in the scope of their search and the exploratory content of their inno-vations (Ahuja & Lampert, 2001; Rosenkopf & Nerkar, 2001). A few studies have examined how organizational design decisions influence ex-ploratory knowledge creation (Jansen, Van Den Bosch, & Volberda, 2006; Sigelkow & Rivkin, 2005). However, with the exception of some qual-itative case study research (Dittrich, Duysters, &
1Burt (1992) also argued structural holes allow actors
freedom from the normative expectations of others in a network, yet research into the influence of network struc-ture on innovation and creativity has stressed informational benefits, rather than control benefits, as the primary causal motor (e.g., Ahuja, 2000; Burt, 2004; Obstfeld, 2005).
de Man, 2007; Gilsing & Noteboom, 2006), re-search has generally ignored the effects of alli-ance network structure and network diversity on exploratory knowledge creation.
The purpose of this study is to address these limitations. I do so by examining the influence of the structure and composition of a firm’s network of horizontal technology alliances on its explor-atory innovation. I focus on horizontal technology alliances for theoretical clarity. Exploratory inno-vation is the creation of technological knowledge that is novel relative to a firm’s extant knowledge stock. Research has often portrayed exploration as a process (March, 1991), yet the manifestation of this process can be observed by examining the explor-atory content of a firm’s innovations (Benner & Tushman, 2002; Rosenkopf & Nerkar, 2001). Ex-ploratory innovations embody knowledge that dif-fers from knowledge used by the firm in prior in-novations and shows the firm has broadened its technical competence (Greve, 2007; Rosenkopf &
Nerkar, 2001).2
Understanding the origins of exploratory innova-tion is an important endeavor. Because the results of exploration (versus exploitation) typically take longer to realize, are more variable, and produce lower average returns, organizations generally pur-sue exploitative innovation at the expense of ex-plorative innovation (March, 1991). They face a fundamental challenge: although exploitation im-proves an organization’s short-term performance, exploration increases its long-term adaptability and survival (Levinthal & March, 1993). This formula-tion does not suggest that exploratory innovaformula-tion is preferred over incremental innovation, only that a balance is necessary (March, 1991). The strong in-centives to pursue exploitation at the expense of exploration raise the question of how and when firms are able to explore effectively. Research has documented the propensity of firms to pursue local search and exploitative innovation (e.g., Dosi, 1988; Helfat, 1994), but much less is known about how and when firms overcome this predisposition and develop exploratory innovations. Explaining the production of exploratory innovations should provide a better understanding of how organiza-tions are able to thrive and survive.
I derive two predictions about the effect of hori-zontal technology alliances on firm exploratory in-novation. First, in highlighting the role of network composition, I draw on the recombinatory search literature (e.g., Fleming, 2001) and examine the benefits and costs of increasing network technolog-ical diversity for exploratory innovation. I predict network diversity has an inverted U-shaped effect on firm exploratory innovation performance. Sec-ond, building on interfirm learning and network research, I argue that the extent to which a firm’s partners are densely interconnected generates trust and reciprocity, which enhance the benefits of net-work diversity and mitigate some of its costs. I predict the density of a firm’s alliance network positively moderates the effect of network diversity on the firm’s exploratory innovation performance. In the empirical work reported here, I tested these predictions on a panel of 77 leading commu-nications equipment manufacturers during 1987– 97 and found partial, yet robust, support for both hypotheses. A positive linear effect of network di-versity emerged, rather than a curvilinear effect, and a positive linear interaction between diversity and density, rather than a curvilinear interaction. This study contributes to the alliance and innova-tion literatures by addressing significant gaps in research on the influence of alliance networks on firm innovation. This is the first study of which I am aware that investigates the influence of alliance network structure and composition on firm explor-atory innovation. The results show the technologi-cal diversity in a firm’s alliance network and the density of the network increase exploratory inno-vation, independently and in combination. The re-sults also suggest the presence of structural holes in a firm’s network is not a necessary condition for providing the firm with access to diverse informa-tion. The extent to which an actor’s network is composed of alters with diverse knowledge bases provides it access to diverse information, indepen-dent of network structure. The benefits of network closure and access to diverse information and know-how can coexist in a firm’s alliance network, and combining the two increases the firm’s explor-atory innovation. Because I find network diversity begets diverse innovations (Kauffman, 1995), the results suggest that dense networks populated by diverse actors generate more, rather than less, di-verse knowledge.
THEORY AND HYPOTHESES
To understand when alliances influence a firm’s exploratory innovation, I build on two complemen-tary theoretical bases: recombinatory search and
2As He and Wong stated, “Exploration versus
exploi-tation should be used with reference to a firm itself and its existing capabilities, resources and processes, not to a competitor or at the industry level” (2004: 485). Thus, “one can only view acts of exploration or exploitation relative to a particular actor’s vantage point” (Adner & Levinthal, 2008: 49).
social capital. The recombinatory search literature casts innovation as a problem-solving process in which solutions to valuable problems are discov-ered via search (Dosi, 1988). Search processes lead-ing to the creation of new knowledge typically in-volve the novel recombination of existing elements of knowledge, problems, or solutions (Fleming, 2001; Nelson & Winter, 1982) or reconfiguring the ways knowledge elements are linked (Henderson & Clark, 1990). Search is uncertain, costly, and guided by prior experience (Dosi, 1988). Over time, feedback from past search efforts becomes embodied in organ-izational routines, which efficiently guide current search efforts (Nelson & Winter, 1982).
Firms create knowledge by engaging in local and distant search (March, 1991). Local search, which is synonymous with exploitation, produces recom-binations of familiar and well-known knowledge elements and is often the preferred mode of search (March, 1991; Stuart & Podolny, 1996). In contrast, distant search, or exploration, involves recombina-tions of novel, unfamiliar knowledge and involves higher costs and uncertainty (March, 1991). Although distant search can be less efficient and less certain than local search, it increases the variance of search and the potential for highly novel recombinations (Fleming, 2001; Levinthal & March, 1981).
Innovation search research has primarily focused on where firms search for solutions (i.e., local ver-sus distant); the interfirm learning literature, on the other hand, has emphasized how firms search. Ac-cording to this research, interfirm relationships are a mechanism for search and a medium of knowl-edge transfer (Ingram, 2002). Because knowlknowl-edge is widely and heterogeneously distributed (von Hayek, 1945), the exchange of knowledge is neces-sary for recombination (Nahapiet & Ghoshal, 1998). Yet the nature of knowledge involved in innovation poses challenges to exchange. Technical innova-tion involves tacit and socially embedded knowl-edge (Dosi, 1988). Technology is knowlknowl-edge em-bedded in communities of practitioners (Layton, 1974) who develop tacit understandings of how to solve problems related to its use and reproduction (von Hippel, 1988). Such knowledge is also stored in organizational routines (Nelson & Winter, 1982). The specialized, tacit, and embedded nature of technical knowledge makes market trading for it subject to severe exchange problems (Teece, 1992). Firms that can identify potentially useful elements of technological knowledge, conceive of how these elements can be fruitfully combined, and effectively access and assimilate this knowledge increase their potential for knowledge creation (Galunic & Rodan, 1998). Strategic alliances are important in each of these aspects of successful recombination.
Strategic alliances are a means of accessing knowledge a firm does not have and can be an effective medium of knowledge transfer and inte-gration (Hamel, 1991). Alliances provide a firm with direct and repeatable access to its partners’ organizational routines, which reduces its ambigu-ity about a partner’s knowledge and increases the efficacy of its transfer and assimilation (Jensen & Szulanski, 2007). Because of the increased social interaction and enhanced incentive alignment and monitoring features they provide, alliances are in-stitutions better suited than market transactions for the repeated exchange of tacit, routine-embedded knowledge (Teece, 1992).
Although alliances provide access to external knowledge, they do not guarantee its effective de-tection, transfer, and assimilation. These processes, and thus the odds of successful recombination, are influenced by the incentives partners have to coop-erate and share knowledge with each other (Hamel, 1991). Because the risk of opportunism is pro-nounced in horizontal technology alliances, effec-tive cooperation and knowledge sharing are diffi-cult to achieve (Gulati & Singh, 1998). Alliance research has typically emphasized the role of for-mal governance mechanisms—such as detailed contracts, the use of equity as a “hostage,” and joint venture structures—in curbing opportunism and increasing cooperation (e.g., Kogut, 1988; Mowery, Oxley, & Silverman, 1996; Sampson, 2007). Other research has suggested that mutual trust and reci-procity norms between partners provide effective and efficient informal governance (Dyer & Singh, 1998; Kale, Singh, & Perlmutter, 2000).
Trust and reciprocity serve as social control mechanisms that mitigate opportunism and safe-guard exchange in alliances (Dyer & Singh, 1998). As such, they are forms of social capital, because they represent resources that are instrumentally valuable for, and appropriable by, partners in a social exchange relationship (Coleman, 1988). The extent to which social capital exists in a firm’s network of alliances can increase the firm’s access to its partners’ knowledge, the motivation of its partners to transfer knowledge, and the efficiency of knowledge exchange and transfer (Inkpen & Tsang, 2005), resulting in more successful recom-binations (Galunic & Rodan, 1998). In the next two sections, I build on the recombinatory search lit-erature and research on interfirm networks and social capital to develop predictions about how and when network technological diversity and network density influence a firm’s exploratory innovation performance.
Network Technological Diversity
Diversity refers to the extent to which a system consists of uniquely different elements, the fre-quency distribution of these elements, and the de-gree of difference among the elements (Stirling, 2007). Thus, I define alliance network technologi-cal diversity as the extent to which the technologies pursued by a firm’s alliance partners are different from one another and from those of the focal firm. Although network diversity provides benefits for a firm’s exploratory innovation efforts, it also poses significant costs. Diversity affects the relative nov-elty of knowledge available in a network and the ease with which a firm can recognize, assimilate, and utilize this knowledge.
Increasing network diversity increases the rela-tive novelty of the knowledge a firm can access. Because exploratory innovations embody relatively novel knowledge, a necessary condition for firm exploratory innovation is access to dissimilar knowledge (Greve, 2007; Jansen et al., 2006). Diver-sity increases the number and variety of possible combinations and the potential for highly novel solutions (Fleming, 2001). The “value of variance” (Mezias & Glynn, 1993) in distant search is that though it increases failures, it also increases the number of highly novel solutions (Levinthal & March, 1981). In contrast, individuals and organi-zations that exploit established competences in their innovative problem-solving efforts typically experience more certain and immediate returns, but produce mostly incrementally innovative solu-tions (Audia & Goncalo, 2007; Dosi, 1988). Search-ing diverse knowledge domains challenges existSearch-ing cognitive structures, including premises and be-liefs about cause-effect relationships (Duncker, 1945), which can promote new associations and lead to highly novel insights and solutions (Simon-ton, 1999). By searching diverse and novel do-mains, firms can develop multiple conceptualiza-tions of problems and soluconceptualiza-tions and apply solutions from one domain to problems in another (Hargadon & Sutton, 1997). Diverse knowledge sources also provide firms with access to diverse problem-solving heuristics (Page, 2007), which can increase the exploratory content of new combina-tions of knowledge (Audia & Goncalo, 2007). Fi-nally, searching diverse, nonredundant knowledge can stimulate intensive experimentation with new combinations, leading to highly novel innovations (Ahuja & Lampert, 2001).
Network diversity also influences a firm’s relative absorptive capacity. As the technological distance be-tween partners increases, their ability to recognize, assimilate, and apply each other’s knowledge
de-clines (Lane & Lubatkin, 1998), increasing the costs of recombinatory innovation (Weitzman, 1998). A firm must expend greater effort and resources to understand and integrate dissimilar knowledge (Cohen & Levinthal, 1990). This can manifest in costly, excessive, and inconclusive experimenta-tion (Ahuja & Lampert, 2001). A firm’s cognitive capacity constraints and its relative inexperience with dissimilar knowledge components will limit its ability to comprehend increasingly complex in-teractions among these components (Fleming & Sorenson, 2001). Moreover, integrating novel knowledge from dissimilar sources often requires changing existing patterns of communication and social exchange, which is difficult in established organizations (Kogut & Zander, 1992). Attempting to assimilate and integrate highly diverse knowl-edge components can lead to information overload, confusion, and diseconomies of scale in innovation efforts (Ahuja & Lampert, 2001). Thus, as a firm’s network diversity increases, its costs of absorbing and utilizing this knowledge greatly increase.
Given these benefits and costs of network diver-sity, I expect it to exhibit a curvilinear effect on a firm’s exploratory innovation. At low levels of di-versity, a firm has a high degree of relative absorp-tive capacity in its portfolio of partners, but the knowledge to which it has access provides little novelty. At high levels of network diversity, ab-sorptive capacity costs are likely to outweigh the benefits of highly novel knowledge. Although in-creasing diversity exponentially increases opportu-nities for novel recombinations (Fleming, 2001), an organization is greatly constrained in its ability to process an abundance of potentially novel recom-binations into usable innovations (Weitzman, 1998). Research has shown that as knowledge com-ponents become more diverse, the chance of their recombination into useful innovations declines, with excessive diversity reducing innovation (Fleming & Sorenson, 2001). In contrast, at a mod-erate level of network diversity a firm’s exploratory innovation efforts benefit from a balance of access to a moderate degree of novel knowledge and mod-erately efficient relative absorptive capacity. Thus, some degree of diversity is valuable for exploratory innovation; too much can be detrimental.
Hypothesis 1. The technological diversity in a firm’s alliance network has an inverted U-shaped relationship with the firm’s subsequent degree of exploratory innovation.
Network Density
Although an alliance provides access to a part-ner’s knowledge, it does not guarantee the effective
detection, transfer, and assimilation of this knowl-edge (Hamel, 1991). The tacit and embedded nature of technological knowledge makes it difficult for partners to detect, transfer, and assimilate (Teece, 1992), reducing its potential for successful recom-bination (Galunic & Rodan, 1998). Increasing net-work diversity worsens this problem, since a firm’s absorptive capacity in relation to its partners will decline (Lane & Lubatkin, 1998). Greater diversity reduces the odds partners share a common under-standing of technical issues, a language for discuss-ing them, and an approach to codifydiscuss-ing knowledge (Cohen & Levinthal, 1990). The exchange hazards in horizontal technology alliances compound these problems. Because partners have incentives to compete, the risk of opportunism is elevated. Such alliances are also inherently uncertain and pose large measurement and monitoring problems (Pisano, 1989). Partners are at risk of involuntary knowledge leakage, the withholding of effort and resources needed to achieve alliance goals, misrep-resentation of newly discovered knowledge, and challenges in transferring tacit knowledge devel-oped during the relationships (Gulati & Singh, 1998). Network diversity also compounds these problems. Increasing diversity increases the rela-tive novelty of knowledge and the variety of tacit knowledge, thereby increasing the amount of unique tacit knowledge. High novelty and tacitness increase partner uncertainty and contractual haz-ards (Pisano, 1989). Technological diversity in-creases coordination problems and the potential for costly contractual renegotiations (Sampson, 2004). These exchange hazards can reduce cooperation and knowledge sharing, hindering a firm’s recom-bination efforts.
The extent to which a firm’s alliance partners are densely interconnected mitigates some of the costs and amplifies some of the benefits of increasing network diversity, thus positively moderating its effect on exploratory innovation. Dense networks facilitate the production of trust and reciprocity among networked firms, which decrease exchange hazards in alliances, increase cooperation among partners, and mitigate absorptive capacity prob-lems. These problems become more challenging, and thus more important to resolve, as network diversity grows.
Network density promotes trust and reciprocity between partners because they share common third-party partners. Dense networks allow firms to learn about current and prospective partners through common third parties, reducing informa-tion asymmetries among firms and increasing their “knowledge-based trust” in one another (Gulati, Nohria, & Zaheer, 2000). Network closure also
pro-motes trust by increasing the costs of opportunism (Coleman, 1988). Because a firm’s behavior is more visible in a dense network, an act of opportunism can damage its reputation, jeopardizing its existing alliances and reducing future alliance opportuni-ties (Gulati, 1998). Because the costs of opportun-ism can outweigh the benefits, firms will refrain from such behavior. Thus, dense networks also generate “enforceable” or “deterrence-based” trust (Kreps, 1990; Raubb & Weesie, 1990). Research has provided empirical support for these arguments (Gulati & Sytch, 2008; Holm, Eriksson, & Johanson, 1999; Husted, 1994; Robinson & Stuart, 2007; Rooks, Raub, Selten, & Tazelaar, 2000; Uzzi, 1996). Network density also generates reciprocity exchanges in which partners share privileged resources because they expect recipients will repay them with some-thing of equivalent value (Coleman, 1988). A firm can encourage reciprocity between two of its ners by transferring reciprocal obligations one part-ner owes to the firm to the other partpart-ner (Uzzi, 1997). Dense networks also promote reciprocity by protecting relationships from opportunism, in-creasing actors’ confidence that obligations for re-payment will eventually be met (Coleman, 1988).
The trust and reciprocity benefits of dense net-works can mitigate some of the exchange hazards and challenges to effective interfirm cooperation associated with greater network diversity. Trust and reciprocity generated by network density act as informal safeguards of dyadic exchange, supple-menting formal alliance governance mechanisms (Powell, 1990). Given the challenges of formal gov-ernance in horizontal technology alliances among technologically diverse firms, informal governance becomes more important in mitigating opportun-ism and promoting cooperation as diversity in-creases. Informal governance reduces the threat of opportunism and increases each partner’s motiva-tion to cooperate and share resources (Dyer & Singh, 1998). Trust reduces the extent to which alliance partners protect knowledge, increases their willingness to share knowledge, and increases in-terfirm learning and knowledge creation (Kale et al., 2000; Larson, 1992). Reciprocity norms rein-force this motivation to share, since firms can be confident partners will reciprocate (Dyer & No-beoka, 2000). As a result, the information and know-how shared will be less distorted, richer, and of higher quality (Dyer & Nobeoka, 2000; Uzzi, 1997). Research has suggested dense interfirm networks are better for transferring and integrating complex and tacit knowledge than networks with structural holes (Dyer & Nobeoka, 2000; Kogut, 2000).
Alliance network density also reduces absorptive capacity problems related to growing network
di-versity. Network closure promotes intense social interaction, experimentation, joint problem solv-ing, and triangulation, which enhance a firm’s abil-ity to absorb and apply increasingly diverse partner knowledge. The trust and reciprocity benefits of
network closure promote intense interaction
among personnel from partnered firms (Larson, 1992), which improves the detection and transfer of tacit and embedded knowledge (Zander & Kogut, 1995). Intense interaction can also lead to the cre-ation of partner-specific knowledge-sharing rou-tines that facilitate knowledge transfer (Lane & Lu-batkin, 1998). The social capital produced in dense alliance networks encourages such relation-spe-cific investments (Walker, Kogut, & Shan, 1997).
The trust and reciprocity benefits of network clo-sure also increase partners’ joint problem-solving efforts and stimulate experimentation with differ-ent knowledge combinations, improving knowl-edge detection and transfer from diverse partners (Dyer & Nobeoka, 2000; Uzzi, 1997). Trust and rec-iprocity can also increase a partner’s motivation to “teach” (Szulanski, 1996), which is more important for student firms as partner diversity increases, since they should find it easier to learn unaided from similar partners (Szulanski, 1996). Alliance partners also provide alternative interpretations of technical problems and solutions, allowing a firm to compare, contrast, and triangulate these perspec-tives (Nonaka, 1994). Alternative perspecperspec-tives dif-fuse rapidly in dense networks (Smith-Doerr & Powell, 2005) and are more valuable when partners are diverse, since a variety of perspectives in-creases the chances some will be useful in a firm’s recombination efforts (Nonaka, 1994). Finally, the rapid flow of information in dense networks pro-vides firms with more opportunities to share and expand their understanding of technical issues and can help establish a shared mode of discourse (Smith-Doerr & Powell, 2005), allowing diverse partners to more efficiently communicate with and learn from one another (Kogut & Zander, 1996).
In sum, increasing network density improves a firm’s ability to absorb and utilize knowledge from more diverse partners. I expect these benefits of network density to moderate the curvilinear effect of network diversity on firm exploratory innova-tion in four distinct ways. First, increasing density will increase the slope (i.e., strength) of the positive relationship between diversity and exploratory in-novation (i.e., the positive slope to the left of the peak of the curve). Second, increasing density will increase the amplitude of the curvilinear effect of diversity. That is, as density increases, the maxi-mum value of exploratory innovation achieved will increase. Third, the value of diversity that
maxi-mizes exploratory innovation will increase as den-sity increases, shifting the peak of the curve to higher values of diversity. Finally, increasing den-sity will reduce the slope of the negative relation-ship between diversity and exploratory innovation. That is, after the effect of diversity turns negative, increasing density will dampen the negative effect of diversity on exploratory innovation.
Hypothesis 2. The density of a firm’s alliance network moderates the curvilinear relationship between network diversity and exploratory in-novation in such a fashion that increasing den-sity will: (a) increase the slope of the positive effect of diversity, (b) increase the amplitude of the effect of diversity, (c) increase the value of diversity that maximizes exploratory inno-vation, and (d) reduce the negative effect of diversity.
METHODOLOGY Sample and Data
The research setting for this study was the global telecommunications equipment industry (SIC 366). Firms in this industry produce and market hard-ware and softhard-ware that enable the transmission, switching, and reception of voice, images, and data over both short and long distances using digital, analog, wire line, and wireless technology. I chose this setting for two reasons. First, during the 1980s and 1990s this industry experienced significant changes in technology and competition, resulting in a growing use of technology alliances by incum-bents (Amesse, Latour, Rebolledo, & Se´guin-Du-lude, 2004). Second, since I used patent data, I chose to study an industry in which firms routinely and systematically patent their inventions (Hage-doorn & Cloodt, 2003; Levin, Klevorick, Nelson, & Winter, 1987).
To minimize survivor bias and right censoring, I limited the study period to 1987–97. I limited the sample frame to public companies to ensure the availability and reliability of financial data. I lim-ited the sample to the firms in the industry with the largest sales because complete and accurate alli-ance data are more available for industry leaders than for smaller firms (Gulati, 1995). To minimize survivor bias, I identified the top-selling firms in the industry at the beginning of the study period rather than the end because numerous mergers, re-structurings, and failures occurred during the study period (Amesse et al., 2004). To minimize the influ-ence of right censoring, I ended the study period in 1997 to allow sufficient time for the (non)approval of patent applications that sample firms made during
the period (also see footnote 4). Following prescrip-tions for establishing network boundaries in empiri-cal research (Laumann, Marsden, & Prensky, 1983), I restricted the network to both firms and alliances that focused on the telecommunications equipment in-dustry. Recent alliance network research has used similar network construction criteria (Rowley, Beh-rens, & Krackhardt, 2000; Schilling & Phelps, 2007). These sampling criteria resulted in a sample of 77 firms headquartered in 13 countries.
I used patent data to measure technological knowledge because patents are valid and robust indicators of knowledge creation (Trajtenberg, 1987). Knowledge is instantiated in inventions, and patents are measures of novel inventions externally validated through the patent examination process (Griliches, 1990). A patent application represents a positive expectation by an inventor of the eco-nomic significance of his or her invention, since getting such protection is costly (Griliches, 1990). Patents measure a codifiable portion of a firm’s technical knowledge, yet they correlate with mea-sures that incorporate tacit knowledge (Brouwer & Kleinknecht, 1999). For these various reasons, pat-ents are a reliable and valid measure of innovation in the telecom equipment industry (Hagedoorn & Cloodt, 2003).
Information on U.S. patents was obtained from Delphion. Using patents from a single country maintains consistency, reliability, and comparabil-ity across firms (Griliches, 1990). U.S. patents are a good data source because of the rigor and proce-dural fairness used in granting them, the large in-centives firms have to obtain patent protection in the world’s largest market for high-tech products, the high quality of services provided by the U.S. Patent and Trademark Office (USPTO), and the rep-utation of the United States for providing effective intellectual property protection (Pavitt, 1988; Riv-ette, 1993). I used the application date to assign a granted patent to a firm because this date closely captures the timing of knowledge creation (Grili-ches, 1990). Because patents are often assigned to subsidiaries, I carefully aggregated patents to the
firm level.3
The collaboration data were obtained from
mul-tiple sources. I initially collected alliance data from the SDC Alliance Database. Although this database provided substantial content, it had many limita-tions. I overcame these limitations through system-atic archival research using annual reports, 10K and 20F filings, Moody’s Manuals, Factiva, Lexis-Nexis, and Dialog. These last three databases index the historical full texts of hundreds of business publications from all regions of the world and in-clude articles translated to English from their orig-inal languages, and non-English publications. I conducted broad keyword searches to identify all instances of interfirm cooperation involving the sample firms. Individuals fluent in the respective language read non-English articles and reports, identified instances of interfirm cooperation, and translated the documents into English. I recorded only collaborations that could be confirmed in mul-tiple sources. Around 1,200 annual reports and Se-curities and Exchange Commission (SEC) filings and over 180,000 electronic articles were exam-ined, and over 8,500 relevant news stories were printed out. Overall, the data set from which this study draws includes 7,904 alliances and 1,967 acquisitions initiated during 1980-96. I reviewed every record from the SDC data and corrected du-plicate entries and other errors and omissions using secondary sources.
Firm attribute data were collected from Compus-tat, annual reports, SEC filings, the Japan Company Handbook, Worldscope, and Global Vantage. Measurement: Dependent Variable
Exploratory innovation. Exploratory innovation
is the creation of technological knowledge by a firm that is novel relative to its existing knowledge stock (Benner & Tushman, 2002; Rosenkopf & Nerkar, 2001). Following prior research (Benner & Tush-man, 2002, Katila & Ahuja, 2002; Rosenkopf & Nerkar, 2001), I measured exploratory innovation using patent citations. I began with the list of U.S. patent classes that corresponded to the telecommu-nications equipment industry at the beginning of the sample period (see Table 1). I assessed the exploratory innovation of firm i in year t by classi-fying and tabulating all citations in the firm’s tele-communications equipment patents applied for in year t (and eventually granted). I traced each cita-tion to determine if the firm had used the same citation or if the citation was to a patent developed by the firm during the seven years before the focal year. I used a seven-year window because organi-zational memory in high-tech firms is imperfect, causing the value of knowledge to depreciate rap-idly over time (Argote, 1999) and creating
signifi-3I identified all divisions, subsidiaries, and joint
ven-tures of each sample firm (using Who Owns Whom and the Directory of Corporate Affiliations) as of 1980. I then traced each firm’s history to account for name changes, division names, divestments, acquisitions, and joint ven-tures and obtained information on the timing of these events. This procedure yielded a master list of entities that I used to identify all patents belonging to sample firms for the period of study.
cant problems for intertemporal knowledge transfer (Nerkar, 2003). Although prior research (e.g., Katila & Ahuja, 2002) has used a 5-year window to assess exploration, I chose a 7-year window because the median age of cited patents in telecom technologies is about 6.5 years (Hicks, Breitzman, Olivastro, & Hamilton, 2001). Using this window, I classified each citation as “new” or “used.” I computed the variable as the result of dividing new citations by
total citations (exploratory innovationsit⫽ new
cita-tionsit/total citationsit.) Because this formula mea-sures a share of new citations, rather than their full count, it captures a firm’s propensity to produce
ex-ploratory innovations, independent of firm scale.4
The extent to which a firm draws on elements of knowledge (e.g., patent citations) it has previously used reflects its practice of local search and exploi-tation of its extant knowledge stock. The extent to which it uses citations with which it has no expe-rience is indicative of distant search and explor-atory innovation (Benner & Tushman, 2002). This measure ranges from pure exploitation (no explo-ration) at the low end to pure exploration (no ex-ploitation) at the high end. It is consistent with research that has conceptualized and measured ex-ploitation and exploration, or local and distant
search, as the ends of a continuum5 (Benner &
4Two aspects of the patent data used to construct this
measure merit discussion. First, during the period of study, the USPTO did not publish patent applications. A patent application date was only observable when a patent was granted. Because I observed patents using their date of application and because there is a delay between the date of application for a patent and its even-tual granting, I may not have observed all patents applied for in a particular year and eventually granted, because the USPTO had not rendered a decision by the time I collected my patent data. The influence of such a right-censoring bias, caused by the delay between patent ap-plication and issuance, is likely to be negligible in this study. Around 99 percent of all applications are re-viewed within five years of application (Hall et al., 2001), which is the period between the end of the sample (1997) and the last year of patent data collection (2002). Second, patent examiners often add citations to patent applications (Alcacer & Gittelman, 2006), which suggests applicant firms are not necessarily aware of all cited patents. Third-party citations often manifest as noise in the measurement of patent-based variables (Jaffe, Trajtenberg, & Fogarty, 2002). Noise in the measurement of a dependent variable in-creases standard errors and reduces the likelihood of find-ing statistically significant effects (Gujarati, 1995).
5Though I focus on one domain of search (i.e.,
tech-nological knowledge), firms search multiple domains, such as customer and geographic space (Gupta, Smith, & Shalley, 2006; Sidhu et al., 2007). Portraying exploitation
TABLE 1
Primary U.S. Patent Classes Used to Represent Telecommunications Equipmenta
Class Number Title
178 Telegraphy
179 (discontinued) Telephony
329 Demodulators
332 Modulators
333 Wave transmission lines and networks
334 Tuners
340 Communications: electrical
341 Coded data generation or conversion 342 Communications: directive radio wave
systems & devices
343 Communications: radio wave antennas
348 Television
358 Facsimile and static presentation processing
359 Optics: systems (including communication) and elements 367 Communications, electrical: acoustic
wave systems and devices 370 Multiplex communications 375 Pulse or digital communications 379 Telephonic communications 381 Electrical audio signal processing
systems and devices 382 Image analysis 385 Optical waveguides 455 Telecommunications
725 Interactive video distribution systems
a
Because patents are classified by technological and func-tional principles, they do not map easily to product-based in-dustrial definitions such as SIC codes (Griliches, 1990). That is, there is not a one-to-one mapping between primary patent classes and industries. Multiple patent classes are used in a single industry, and a single patent class can be used in multiple industries. Consequently, to identify the areas of technology that constitute telecommunications equipment, I needed to develop a concordance between primary patent classes and the three-digit SIC code 366, “communications equipment.” To do so, I utilized both Silverman’s (1996) concordance method and con-cordances provided by experts. I used the concordance for com-munications equipment developed by scholars at Science and Technology Policy Research (SPRU), a unit of Sussex University in the United Kingdom, and the concordance developed by the Community of Science Inc., an internet company that provides collaborative tools and services for research scientists and engi-neers. I identified the primary patent classes common to both of these expert-based concordances as a baseline and then com-pared this list of classes with a rank-ordered list delineating the degree to which specific international patent classes (IPCs) were associated with SIC 366 as the industry of manufacture as of 1988. To make this comparison, I used the USPTO’s USPC-IPC concordance. The primary classes listed in the baseline concor-dance were associated with the highest ranked IPC classes as-sociated with U.S. SIC 366 (except for class 725, which did not exist in the late 1980s). This indicated that the 22 primary classes used in this study to represent communications equip-ment technology in this table are most frequently associated with SIC 366.
Tushman, 2002; Greve, 2007; Sidhu, Commandeur, & Volberda, 2007).
As a robustness check, I applied an alternative measure of exploratory innovation from prior re-search (e.g., Ahuja & Lampert, 2001; McGrath & Nerkar, 2004). I computed this measure as the num-ber of new three-digit technology classes in which firm i patented in year t, classifying a technology class as new if the firm had not patented in that class in the past seven years. The USPTO assigns patents to about 450 technology classes, with each class demarcating an area of technology. The extent to which a firm enters new technological domains is indicative of exploration (Ahuja & Lampert, 2001; McGrath & Nerkar, 2004). This measure was broader than the citation-based measure since it took into account all technology classes in which a firm might patent.
Measurement: Explanatory Variables
Following prior research (e.g., Ahuja, 2000; Stu-art, 2000), I sampled alliances involving technology development or exchange because my phenome-non of interest and theory concerned the transfer and creation of technological knowledge. I ex-cluded unilateral licensing deals and alliances formed for the sole purpose of marketing, distribu-tion, or manufacturing.
Network technological diversity. To measure
network technological diversity, I employed Rodan and Galunic’s (2004) measure of knowledge hetero-geneity. This measure incorporates information about the knowledge distance between a focal actor and each of its partners and the distances among the part-ners. I began at the dyad level and measured the technological distance between pairs of firms using Jaffe’s (1986) index. For each firm-year, I measured the distribution of a firm’s patents across primary patent classes. Following Sampson (2007), I used a moving four-year window to establish a firm’s patent-ing profile. This distribution located a firm in a mul-tidimensional technology space, captured by a
K-di-mensional vector (fi⫽ [fi1. . . fik], where fikrepresents
the fraction of firm i’s patents that are in patent class k). This approach rests on an assumption that the distribution of a firm’s patents across classes reflects the distribution of its technical knowledge
(Jaffe, 1986). The technological distance, d, be-tween firms i and j in year t was calculated as:
dijt⫽ 1 ⫺
冋
冘
k⫽ 1 K fikfjk冒
冉
冘
k⫽ 1 K fik2冊
1/ 2冉
k冘
⫽ 1 K fjk2冊
1/ 2册
.This measure was bounded between 0 (complete similarity) and 1 (maximum diversity) and sym-metric for the two firms. I used these pairwise distance values to construct annual distance
matri-ces, Dt, which reflected the technological distances
between all possible pairs of sample firms.
Next, I computed the uniqueness of the knowl-edge of each partner j in firm i’s alliance network in year t. The uniqueness of firm j is a function of the uniqueness of its partners, k, and firm j’s distance from them. Following Rodan and Galunic (2004), I
defined the uniqueness of firm j, uj, as:
uj⫽
冘
k
djk⫻ uk.
The uniqueness of each firm is found in the solu-tion of the eigen equasolu-tion(U ⫽ DU), where U is an
eigenvector of D and is its associated eigenvalue.
The elements of U are the uniqueness values for each firm, and D is the matrix of pairwise techno-logical distances. I measured the technotechno-logical di-versity available to firm i in its (ego) network of alliance partners in year t as:
Network technological diversityit⫽ 1 N
冘
j⫽ 1
N dijuj,
where dijis partner j’s distance from i andujis j’s
uniqueness score computed for i’s N partners. The 1/N term compensates for the fact that lambda in-creases linearly with network size. This measure increases linearly with the distances among i and its partners (Rodan & Galunic, 2004).
Network density. To measure ego network
den-sity, I constructed annual adjacency matrices for the period 1987–96 that indicated the presence of a technology alliance, in existence at the end of a focal year, between all possible undirected pair-wise combinations of sample firms. An alliance with more than 2 firms entered the adjacency ma-trix as separate dyadic combinations of all firms in the alliance. Of all sample alliances, 89 percent involved only 2 firms, and the average alliance had 2.38 firms. Because alliances often endure longer than one year, constructing adjacency matrices us-ing only alliances formed in a focal year would have understated the true connectivity of the net-work. Consequently, I collected alliance data for each firm beginning in 1980 and researched each
and exploration as ends of a continuum in one domain of search does not preclude the possibility that firms can simultaneously achieve high levels of both exploitation and exploration in multiple domains (Gupta et al., 2006). A universal argument about the mutual exclusivity or independence of exploitation and exploration may be impossible (Gupta et al., 2006).
alliance to identify its date of dissolution or
con-tinuance through the last sample year.6
Ego network density was the percentage of all possible ties among an ego’s alters that had been formed (Scott, 1991). Ego networks in which a firm’s alliance partners are themselves allied imply higher values of density. To test the robustness of the effect of density, I substituted Burt’s (1992) measures of efficiency and then constraint into al-ternative specifications. The Appendix presents these specifications. Both efficiency and constraint are measures of triadic closure (see Borgatti [1997] for a comparison). Figure 1 presents an example of a sample firm’s ego network, specifically, Motoro-la’s network of technology alliances at the end of 1992, and lists the values for the density, effi-ciency, and constraint of this network. Algebraic explanations of each measure are also shown. Control Variables
To minimize alternative explanations and isolate the marginal effects of the explanatory variables, I controlled for several firm- and alliance-level vari-ables whose influence on exploratory innovation might be confounded with the explanatory vari-ables. Given the firm-level analysis used in this study, I aggregated alliance-level observations to the firm level. I used multiple-year moving win-dows of differing lengths to compute five control variables. These window lengths ranged from four to seven years and differed by control variable. I based the choice of window length for each control variable on prior research. Using alternative
win-dow lengths (⫾1 year) for these control variables
did not substantively change the results of the ex-planatory variables presented in Table 2.
Network size. More alliance partners may
pro-vide a firm with access to greater technical diver-sity. Moreover, measures of ego network density are sensitive to network size, making network size an important control variable (Friedkin, 1981). I computed network size as the natural logarithm of the number of telecom technology alliance partners maintained by firm i in year t.
Alliance duration. Alliance longevity can lead to
greater interfirm trust (Gulati, 1995), stronger reci-procity norms (Larson, 1992) and relation-specific routines (Levinthal & Fichman, 1988), increasing interfirm learning (Simontin, 1999). I measured al-liance duration as the average number of years firm i had participated in its existing telecom technol-ogy alliances at the end of year t (see footnote 6).
Repeated ties. Prior ties between firms can
in-crease interfirm trust (Gulati, 1995), the develop-ment of relation-specific learning heuristics, and interfirm learning (Lane & Lubatkin, 1998). Follow-ing Gulati and Gargiulo (1999), I calculated re-peated ties as the average number of alliances firm i had formed with its current group of alliance partners in the five years prior to year t.
Joint venture. Research has suggested equity
joint ventures are superior governance mechanisms for interfirm learning and knowledge transfer (Kogut, 1988; Mowery et al., 1996). I computed the variable as the proportion of firm i’s telecom tech-nology alliances governed by equity joint ventures in year t.
International alliance. International alliances
provide access to diverse knowledge (Rosenkopf & Almeida, 2003), but they experience greater coor-dination and communication problems and cul-tural conflicts than domestic alliances, and this experience diminishes interfirm learning (Lyles & Salk, 1996). I measured this variable as the fraction of firm i’s telecom technology alliances in year t involving foreign firms.
Partners’ market overlap. Because partners
tend to protect their knowledge when they are product-market competitors, overlaps in partners’ markets can impede interfirm knowledge transfer (Dutta & Weiss, 1997). I computed market overlap as the proportion of firm i’s portfolio of telecom technology alliances in year t having partners with the same primary four-digit SIC code as firm i.
Firm sales. Firm size can have both negative and
positive effects on firm innovation (Teece, 1992). I controlled for firm size using the natural log of sales (in millions of U.S. dollars) for firm i in year t.
6I researched each alliance using the sources
de-scribed previously. I also contacted company personnel to identify dissolution dates, which proved very useful in identifying the termination or ongoing status of joint ventures (JVs). For nearly all JVs, I was able to identify the months they were ended or their ongoing status at the end of the sample period. For each remaining JV, I as-sumed it existed until the end of the last year in which it was documented or until the end of the year after the year it was founded, whichever was later. For non-JV alli-ances, I recorded termination on the basis of specified tenure, if mentioned in the archival sources, or an-nouncement of dissolution (either from archival sources or company contact). In cases in which I could not es-tablish precise dissolution, I followed Ahuja (2000) and presumed an alliance to exist until the end of the last year in which it was documented or until the end of the year after the year it was founded, whichever was later. I performed a t-test of the difference in mean duration between alliances with formal dissolution announce-ments and those with assumed dissolution dates and found no significant difference.
FIGURE 1
Motorola’s 1992 Ego Network Structure of Technology Alliancesa
a
In the figure, Motorola is the focal actor, or ego. Below are the values of ego network density, efficiency and constraint for Motorola’s 1992 technology alliance network and an explanation of each measure. Burt (1992) provides a detailed explanation of the measures of efficiency and constraint and Borgatti (1997) provides a comparison of the three measures.
The values for the density, efficiency, and constraint of this network and their algebraic computation are as follows:
Density⫽ 26.67%
Ego network densityi⫽
冋冉冘
j冘
qxjq冊
冒
冉再
N冉
N⫺ 1冊冎
冒
2冊册
⫻ 100, j ⫽ q,where xjqrepresents the relative strength of the tie between alter j and alter q, and N represents the number of alters to which ego i is connected.
Because I treated alliances as either present or absent (i.e., they do not vary in terms of strength), all values of xjqwere set to 1 if a relationship
existed and 0 otherwise. The term [N(N⫺ 1)] was divided by 2 to reflect that alliances are undirected ties. Variable range, 0–100%.
Efficiency⫽ 0.75
Ego network efficiencyi⫽
冋冘
j冉
1⫺冘
qpiqmiq冊册
冒
N, j⫽ q,where piqis the proportion of i’s ties invested in the relationship with q, mjqis the marginal strength of the relationship between alter j and alter
q (as I used binary data, all values of mjqwere set to 1 if a tie existed and 0 otherwise), and N represented the number of alliance partners to which
focal firm was connected. This measure could vary from 0 to 1, with higher values indicative of greater efficiency (i.e., structural holes).
Constraint⫽ 0.15
Ego network constrainti⫽
冘
j冋
pij⫹冘
qpiqpqj册
2
, q⫽ i, j,
where pijis the proportion of i’s ties invested in the relationship with j, piqis the proportion of i’s ties invested in the relationship with
q, and pqjis the proportional strength of alter q’s relationship with alter j. This measure can vary from of 0 to 1, with higher values
TABLE 2 Descriptive Statistics and Correlations a Variables Mean s.d. Min. Max. 123456789 1 0 1 1 1 2 1 3 1 4 1 5 1 6 1 7 1. Exploratory innovation 0.72 0.19 0 1 4 2. Network technological diversity 0.22 0.33 0 1.71 .19 3. Network density 33.88 32.99 0 100 .18 ⫺ .27 4. Network size b 1.33 0.99 0 3.56 ⫺ .25 .79 ⫺ .38 5. Alliance duration 3.10 1.87 0 1 4 .08 .14 ⫺ .14 .20 6. Repeated ties 0.38 0.46 0 4 .19 .49 ⫺ .17 .43 .13 7. Joint venture 0.18 0.19 0 1 .02 .39 ⫺ .18 .38 .26 .20 8. International alliance 0.30 0.25 0 1 ⫺ .06 .16 ⫺ .14 .18 ⫺ .02 .08 .28 9. Partners’ market overlap 0.33 0.24 0 1 .02 .04 .08 ⫺ .07 ⫺ .02 ⫺ .09 ⫺ .12 .14 10. Sales b 6.47 2.66 ⫺ 2.99 11.34 ⫺ .17 .64 ⫺ .20 .61 .39 .55 .40 .17 ⫺ .05 11. Firm current ratio 2.32 1.58 0.003 23.46 .13 ⫺ .21 .09 ⫺ .22 ⫺ .19 ⫺ .26 ⫺ .29 ⫺ .17 .06 ⫺ .37 12. Firm R&D intensity 0.13 0.76 0 20.25 .03 .04 .07 .01 ⫺ .27 ⫺ .03 ⫺ .18 ⫺ .08 ⫺ .02 ⫺ .20 .09 13. Firm patent stock 582.06 1,267.74 1 6,875 ⫺ .14 .50 ⫺ .21 .44 .21 .67 .29 .15 ⫺ .11 .63 ⫺ .25 ⫺ .04 14. Firm age 45.39 36.03 2 150 .13 .58 ⫺ .18 .52 .38 .37 .50 .25 ⫺ .04 .74 ⫺ .30 ⫺ .09 ⫺ .11 15. Firm alliance experience 0.14 0.75 0.001 19.25 ⫺ .02 ⫺ .09 .14 ⫺ .18 ⫺ .08 .03 ⫺ .15 ⫺ .07 ⫺ .08 ⫺ .29 .05 .36 .52 ⫺ .16 16. Firm technological diversity 0.76 0.31 0 1 ⫺ .11 .34 ⫺ .19 .32 .32 .25 .40 .11 ⫺ .09 .59 ⫺ .38 ⫺ .02 ⫺ .07 .53 ⫺ .18 17. Firm acquisitions 0.96 1.71 0 1 6 ⫺ .06 .34 ⫺ .08 .32 .03 .07 .24 .06 .34 ⫺ .11 ⫺ .04 ⫺ .06 .38 ⫺ .09 .20 18. U.S.-Canada 0.65 0.48 0 1 ⫺ .1 ⫺ .50 .14 ⫺ .31 ⫺ .30 ⫺ .39 ⫺ .52 ⫺ .29 ⫺ .11 ⫺ .65 .29 .06 .09 ⫺ .61 .13 ⫺ .43 ⫺ .19 19. Europe 0.19 0.39 0 1 .04 .51 ⫺ .10 .39 .14 .06 .50 .21 .12 .18 ⫺ .18 ⫺ .03 ⫺ .06 .55 ⫺ .08 .30 .40 ⫺ an (firms) ⫽ 77; n (observations) ⫽ 707. All correlations greater than 兩.07 兩 are significant at p ⬍ .05. bLogarithm.
Firm current ratio. The availability of slack
resources can increase exploratory search (Singh, 1986) and lead to greater innovative performance (Nohria & Gulati, 1996). I controlled for the un-absorbed slack resources of firm i in year t using its current ratio (current assets/current liabilities) (Singh, 1986).
Firm R&D intensity. A firm’s R&D expenditures
are investments in knowledge creation (Griliches, 1990) and contribute to its ability to absorb extra-mural knowledge (Cohen & Levinthal, 1990). I mea-sured R&D intensity by dividing firm i’s R&D ex-penses by its sales in year t.
Firm patent stock. The more patents a firm has,
the more patents and references it can cite; a large patent stock could thus negatively affect the pri-mary measure of exploratory innovation here. A firm’s patent stock also reflects the depth of its technological resources and absorptive capacity (Silverman, 1999). I controlled for the number of firm i’s patents obtained in the four years prior to the end of year t.
Firm age. As firms age, they tend to exploit their
existing technological competencies rather than ex-plore new and unfamiliar technologies (Sorensen & Stuart, 2000). I operationalized firm age as the number of years from the date of founding of firm i to year t.
Firm alliance experience. Alliance experience
enhances the collaborative capability of a firm, which facilitates interfirm knowledge transfer (Sampson, 2005). I controlled for the number of all types of alliances formed by firm i in the seven years before year t, divided by its sales in year t.
Firm technological diversity. Technologically
diverse firms may be more innovative because of diverse internal knowledge flows (Garcia-Vega, 2006), and they may be more able to absorb external knowledge (Cohen & Levinthal, 1990). I measured firm i’s diversity in year t using a modified Herfin-dahl index (Hall, 2002):
Technological diversityit⫽
冋
1⫺冘
j⫽ 1 J冉
Njit Nit冊
2册
⫻ Nit Nit⫺ 1 ,where Nitis the number of patents obtained by firm
i in the past four years. Njitis the number of patents in technology class j in firm i’s four-year patent stock. This variable could range from 0 to 1 (max-imum diversity).
Firm acquisitions. Acquisitions can enhance
ac-quirer innovation (Ahuja & Katila, 2001). Telecom
equipment firms often use both acquisitions and alliances to source knowledge (Amesse et al., 2004). I controlled for the number of telecom equip-ment acquisitions (i.e., those in which the target company’s primary SIC code was 366) made by firm i during the four years prior to and including year t.
U.S.-Canada/Europe/Asia. I used dummies
de-noting the regional origin of a firm to control for regional effects. “U.S.-Canada” was coded 1 if a firm was headquartered in the United States or Canada. “Europe” was coded 1 if the firm was headquartered in Europe. Asia was the omitted category.
Model Specification and Estimation
The dependent variable was a proportion and presented several challenges to linear regression (Gujarati, 1995). Thus, I used three alternative mod-eling approaches. First, I estimated the models with exploratory innovation as the dependent variable using panel linear regression and robust standard errors. Following common econometric practice (Greene, 1997), I also estimated models with a
log-odds transformation of exploratory innovation.7
Fi-nally, I estimated models using a generalized esti-mating equation (GEE) approach in which I specified a probit link function and an exchange-able correlation matrix and computed robust errors (Papke & Wooldridge, 2005). As a robustness check, I compared the results from these alternative spec-ifications. I included year dummies to control for period effects, such as differences in macroeco-nomic conditions or industry technological oppor-tunity. Either firm-specific fixed or random effects can be used to control for unobserved firm hetero-geneity (Greene, 1997), such as differences in mo-tivations to pursue, and abilities to develop, explor-atory innovations. Because the use of random effects relies on an assumption that errors and re-gressors are uncorrelated, I used a Hausman (1978) test to choose between fixed and random effects. I also checked for first-order serial autocorrelation in the errors. I lagged all independent variables one year, which reduced concerns of reverse causality and avoided simultaneity.
7The transformed variable is as follows:
ln(explor-atory innovation/1 – explorln(explor-atory innovation). Because
the transformation is undefined when exploratory inno-vation is equal to 0 or 1, I recoded these values as follows: 0⫽ 0.0001 and 1 ⫽ 0.9999.
RESULTS
Table 2 reports descriptive statistics and correla-tions. The panel was unbalanced and consisted of 77 firms and 707 firm-year observations. Table 3 presents the results of the panel regression analysis used to test the hypotheses. I report the results for untransformed exploratory innovation for ease of interpretation. The results using a logit transforma-tion and those from GEE estimatransforma-tion are consistent with those reported in Table 3. I estimated models 1–7 using firm random effects for three reasons: (1) significant unobserved heterogeneity was present, (2) Hausman specification tests were not signifi-cant, supporting the use of random effects, and (3) significant serial correlation was not present. Hu-ber-White (or “sandwich”) robust standard errors
are reported, and all significance levels are for two-tailed tests. Multicollinearity does not seem to have unduly influenced the regression results because the average variance inflation factor (VIF) for each model and the VIFs for all variables were below the rule-of-thumb value of ten (Gujarati, 1995).
Hypothesis 1 predicts an inverted U-shaped ef-fect of network technological diversity on firm ex-ploratory innovation. Models 2– 6 in Table 3 pro-vide partial support for this hypothesis. In each of
these models, network technological diversityit⫺1
exhibited a positive and significant effect on ex-ploratory innovation. However, the squared term was not significant in any model in which it was entered. Thus, although I found evidence of a pos-itive linear effect of network diversity, I did not TABLE 3
Results of Random-Effects Panel Linear Regression Analysis Predicting Firm Exploratory Innovationa
Variables Model 1 Model 2 Model 3 Model 4 Model 5 Model 6
Constant 0.85** (0.06) 0.89** (0.07) 0.89** (0.07) 0.89** (0.07) 0.86** (0.07) 0.85** (0.07) Network sizeb ⫺0.03** (0.01) ⫺0.05** (0.01) ⫺0.05** (0.01) ⫺0.04** (0.01) ⫺0.04** (0.01) ⫺0.04** (0.01) Alliance duration ⫺0.01 (0.01) ⫺0.01 (0.01) ⫺0.01 (0.01) ⫺0.01 (0.01) ⫺0.01 (0.01) ⫺0.01 (0.01) Repeated ties 0.01 (0.02) 0.01 (0.02) 0.02 (0.02) 0.01 (0.02) 0.02 (0.02) 0.01 (0.02) Joint venture 0.08 (0.06) 0.08 (0.06) 0.08 (0.06) 0.07 (0.06) 0.05 (0.06) 0.05 (0.06) International alliance ⫺0.09* (0.04) ⫺0.08* (0.04) ⫺0.08* (0.04) ⫺0.08* (0.04) ⫺0.07* (0.04) ⫺0.08* (0.04) Partners’ market overlap ⫺0.04 (0.04) ⫺0.03 (0.04) ⫺0.03 (0.04) ⫺0.02 (0.04) ⫺0.02 (0.04) ⫺0.01 (0.04) Firm salesb ⫺0.02* (0.01) ⫺0.03* (0.01) ⫺0.03* (0.01) ⫺0.02* (0.01) 0.00 (0.01) 0.00 (0.01)
Firm current ratio 0.00 (0.01) 0.00 (0.01) 0.00 (0.01) 0.00 (0.01) 0.00 (0.01) 0.00 (0.01) Firm R&D intensity ⫺0.08 (0.08) ⫺0.08 (0.08) ⫺0.08 (0.08) ⫺0.08 (0.08) ⫺0.07 (0.08) ⫺0.06 (0.08) Firm patent stock/1,000 ⫺0.004** (0.00) ⫺0.002** (0.00) ⫺0.002** (0.00) ⫺0.002** (0.00) ⫺0.003** (0.00) ⫺0.003** (0.00) Firm age 0.00 (0.00) 0.00 (0.00) 0.00 (0.00) 0.00 (0.00) 0.00 (0.00) 0.00 (0.00) Firm alliance experience ⫺0.01 (0.03) ⫺0.02 (0.03) ⫺0.02 (0.03) ⫺0.02 (0.03) ⫺0.02 (0.03) ⫺0.02 (0.03) Firm technological diversity ⫺0.05 (0.04) ⫺0.05 (0.04) ⫺0.05 (0.04) ⫺0.05 (0.04) ⫺0.05 (0.04) ⫺0.05 (0.04) Firm acquisitions 0.00 (0.00) 0.00 (0.00) 0.00 (0.00) 0.00 (0.00) 0.00 (0.00) 0.00 (0.00) U.S.-Canada ⫺0.01 (0.02) ⫺0.01 (0.02) ⫺0.01 (0.02) ⫺0.01 (0.02) ⫺0.01 (0.02) ⫺0.01 (0.02) Europe 0.01 (0.03) 0.00 (0.03) 0.00 (0.03) 0.00 (0.03) 0.01 (0.03) 0.00 (0.03) Network technological diversity 0.06* (0.03) 0.06* (0.03) 0.06† (0.03) 0.07* (0.04) 0.08* (0.04) Network technological diversity squared 0.00 (0.04) 0.00 (0.04) 0.05 (0.04) 0.06 (0.09) Network density 0.05* (0.02) 0.10* (0.04) 0.07 (0.05) Network technological diversity⫻ density 0.46** (0.13) 0.53** (0.14) Network technological diversity squared⫻ density 0.48 (0.36)
Year dummies included Yes Yes Yes Yes Yes Yes
R2 0.13 0.14 0.14 0.16 0.17 0.18
Wald2(df) 4.67** (1) 4.66* (2) 6.18* (3) 9.94** (4) 11.71** (5) an(firms)⫽ 77; n(observations) ⫽ 707. Huber-White robust standard errors are in parentheses.
bLogarithm. †p⬍ .10
* p⬍ .05 ** p⬍ .01 Two-tailed tests.
